Research map / Label-efficient and robust learning / Knowledge distillation
Cross-Architecture Distillation Made Simple with Redundancy Suppression
Weijia Zhang, Yuehao Liu, Wu Ran, Chao Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add decorrelation (off-diagonal) objective to invariance objective (ablation) | RSD · CIFAR-100 (Swin-T to ResNet18) | accuracy: 80.65 -> 83.92 | Improved |
| add decorrelation objective to invariance objective (ablation) | RSD · CIFAR-100 (ConvNeXt-T to ResMLP-S12) | accuracy: 83.40 -> 84.21 | Improved |
| add RSD invariance objective over no distillation (ablation) | student · CIFAR-100 (Swin-T to ResNet18) | accuracy: 74.01 -> 80.65 | Improved |
| add AAD decoupler module (ablation) | RSD · CIFAR-100 (ViT-S to ResMLP-S12) | accuracy: 82.26 -> 82.94 | Improved |
| add AAD decoupler module (ablation) | RSD · ImageNet-1k (ConvNeXt-T to Mixer-B/16) | accuracy: 79.93 -> 80.73 | Improved |
| use RSD loss on logits instead of OFA logit loss (ablation) | logit distillation · CIFAR-100 (Mixer-B/16 to DeiT-T) | accuracy: 70.69 -> 77.22 | Improved |
| use RSD loss on logits instead of KD loss (ablation) | logit distillation · CIFAR-100 (Swin-T to ResNet18) | accuracy: 78.74 -> 83.23 | Improved |
| replace OFA losses with RSD loss (OFA+RSD) (ablation) | OFA · CIFAR-100 (ConvNeXt-T to ResMLP-S12) | accuracy: 81.22 -> 84.46 | Improved |
| replace OFA losses with RSD loss (OFA+RSD) (ablation) | OFA · ImageNet-1k (DeiT-T to ResNet18) | accuracy: 70.64 -> 71.21 | Improved |
| add RSD over OFA-projected intermediate features at stages {1,2,3,4} (ablation) | RSD · CIFAR-100 (Swin-T to ResNet18) | accuracy: 83.92 -> 85.45 | Improved |
| add RSD over OFA-projected intermediate features at stages {3,4} (ablation) | RSD · CIFAR-100 (Mixer-B/16 to DeiT-T) | accuracy: 76.45 -> 76.37 | No clear effect |
| use RSD instead of OFA in cross-architecture distillation | student · CIFAR-100 (ConvNeXt-T to DeiT-T) | accuracy: 75.76 -> 82.46 | Improved |
Closest papers by meaning
- One-for-All: Bridge the Gap Between Heterogeneous Architectures in Knowledge Distillation (NeurIPS 2023 · similarity 0.90)
- Fuse Before Transfer: Knowledge Fusion for Heterogeneous Distillation (ICCV 2025 · similarity 0.88)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.86)
- Distilling Knowledge from Heterogeneous Architectures for Semantic Segmentation (AAAI 2025 · similarity 0.85)
- Understanding the Role of the Projector in Knowledge Distillation (AAAI 2024 · similarity 0.85)
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
